Reflection open-sources Beam, a 501B coding model under Apache 2.0
Reflection has released the weights for Beam, its first model, on Hugging Face under the Apache 2.0 licence. It is a sparse mixture-of-experts model with 501B total parameters and 23B active, built for coding, reasoning and long-horizon agent work. Weights ship in BF16, FP8 and NVFP4 with a speculative-decoding drafter, a safety technical report and Mirror CLI, Reflection's own open-source coding harness. Day-one support covers Transformers, vLLM and SGLang, with more than 40 launch partners. Coming a day after Meta's Muse Spark Open, which uses a limited commercial licence, Beam is the more permissive option for anyone who wants to self-host a frontier-scale US model without licence fine print. Reflection's announcement
What else matters
DeepSeek ships V4.1-Flash-1017
DeepSeek calls its new checkpoint a Pareto improvement over September's V4.1-Flash, beating it on all 11 reported benchmarks by 2.0 to 11.1 points. It scores 77.0 on DeepSWE v1.1 against 74.0 for Claude Opus 5, but trails clearly on Humanity's Last Exam without tools (42.6 vs 56.3). DeepSeek API updates
NVIDIA widens Nemotron 3.5 for agents and desktops
NVIDIA released Nemotron 3.5 Super VL, an open 120B-parameter vision-language model aimed at specialised always-on agents, alongside smaller Bolt models sized for local hardware, down to a 4GB edge variant. It is NVIDIA's pitch for running agent workloads from data centre to desk. NVIDIA blog
NVIDIA pledges $250m of compute to UK agent safety research
NVIDIA announced a five-year compute donation valued at $250m to strengthen the UK's agentic safety and security research, plus partnerships across finance, energy and healthcare. UK AI Minister Kanishka Narayan called it a vote of confidence in the country's research base. AI Magazine
